AI unit economics turns a pile of token counts into a cost per outcome, signifying what it costs to resolve a support ticket, review a code change or summarize a document. Finance does not budget in tokens. It budgets in work the business gets done, so that is the unit worth measuring.
This post explains how to get from tokens to cost per outcome, how showback and chargeback differ, why averages mislead, and what to put in front of finance.
What is AI unit economics?
AI unit economics is the practice of measuring what AI costs per unit of business output, such as per ticket resolved, per code review or per document summarized. Tokens are the input to that calculation. The cost per outcome is the answer.
Think of a fuel-card statement. Finance does not care how many liters were bought. It cares what each delivery cost. A statement that lists only liters tells you the fleet is busy and nothing about whether the business is efficient. AI works the same way: a total of 40 million tokens says very little, while “this assistant costs us a few cents per resolved ticket” is something a manager can act on.
That shift matters for three reasons:
- It compares like with like. Two tools that use very different volumes of tokens can still be compared on what they cost per outcome.
- It connects spend to value. A cost per outcome can be set against what the outcome is worth, which is the basis of any AI ROI conversation.
- It gives budget owners a number they recognize. Cost per query and cost per user are units that finance and department heads already use for other tools.
Additional Read: What is AI tokenomics (and why it isn’t about crypto)
What is the difference between showback and chargeback?
Showback shows each team what its AI usage cost without moving any money. Chargeback goes a step further and bills that cost to the team’s own budget or cost center. Both depend on AI cost allocation, which means knowing who consumed what.
| Showback | Chargeback | |
|---|---|---|
| What happens | Teams receive a report of their AI cost | The cost is moved to the team’s cost center |
| Best when | AI use is new, or attribution is still being validated | Attribution is trusted and departments own their AI budgets |
| Main benefit | Builds awareness with no budget friction | Creates direct accountability for spend |
| Main risk | Reports can be ignored because nothing is at stake | Disputes if the numbers are not trusted |
| What it needs | Usage attributed to a team | Attributed usage, a price per token and finance sign-off |
A common path is to start with showback. It lets teams see their numbers, question them and fix gaps in attribution before any money is involved. Chargeback becomes realistic once everyone agrees the numbers are right.
How should you attribute AI costs?
Attribute AI costs along three dimensions: the user, the project and the department. Each answers a different question, and together they let you move from a provider invoice to a cost center.
- User. Who is consuming? This shows usage patterns, highlights heavy users who may need guidance, and gives you cost per user.
- Project. What is it being used for? This is how you compare initiatives and calculate AI ROI for a specific piece of work.
- Department. Who pays? This maps consumption to a cost center, which is what budgets and chargeback run on.
The three nest naturally: users belong to projects, and projects belong to departments. If you can only capture one at the start, capture the one your finance team already uses to allocate other costs, which is usually the department or cost center.
Attribution has to happen where the request is made. A provider invoice shows totals per model or key, and it cannot be reverse-engineered into users and projects afterwards. That is why teams that want reliable AI cost allocation capture the user, project and department as each request passes through, not at the end of the month.
Why do averages hide the real picture?
An average cost per user can look healthy while one or two heavy users drive most of the spend. Token usage is rarely evenly spread, so the mean often describes nobody in particular.
Here is a hypothetical example. A team of ten people uses an AI assistant for a month. Nine of them use 100,000 tokens each. One uses 1,000,000 tokens, perhaps because of an automated script.
- Total: 1,900,000 tokens.
- Average per user: 190,000 tokens.
- Typical user: 100,000 tokens.
- The one heavy user: more than half of all tokens.
Reporting the average of 190,000 would suggest every user is nearly twice as costly as the typical user really is, and it would hide the single source of the cost. The useful questions are about the spread: who is in the top tier, what are they doing, and is it valuable or wasteful? Showing the top users, a median and a few percentiles alongside the mean makes the picture honest, and it tells you where to look first when you want to reduce costs.
Additional Read: LLM token pricing: Why AI bills are hard to predict
How do you present AI cost to finance?
Present AI cost to finance as a small set of unit metrics, each with a clear formula, an owner and a trend over time. Finance teams trust numbers they can recalculate, so define every metric once and use the same definition in every report.
Here is a simple template. All figures are hypothetical and exist only to show the arithmetic.
| Metric | How to calculate it | Hypothetical example |
|---|---|---|
| Cost per outcome | AI spend for a use case ÷ outcomes completed | $2,000 ÷ 8,000 tickets resolved = $0.25 |
| Cost per user | AI spend ÷ users active in the period | $3,000 ÷ 150 active users = $20 |
| Cost per query | AI spend ÷ number of requests, often quoted per 1,000 | $400 ÷ 100 thousand queries = $4 per 1,000 |
| Share of spend by cost center | One cost center’s AI spend ÷ total AI spend | $2,000 ÷ $5,000 = 40% |
| Net value per outcome | Value of one outcome − cost per outcome | $4.00 − $0.25 = $3.75, if a resolved ticket is worth $4.00 |
A few habits make the template work:
- Start with two or three metrics. Cost per outcome and share of spend by cost center are enough for a first conversation.
- Show the trend. A cost per outcome that is falling month by month is a stronger story than a single figure.
- Name the assumptions. If the value of an outcome is an estimate, say so, and who agreed it.
- Pair averages with the spread. As the previous section showed, a mean alone can mislead.
What does this look like in practice?
In practice, AI unit economics needs three ingredients: usage tagged with a user, project and department at the moment of the request, a price per token applied to that usage, and a way to roll the result up to cost centers. A gateway or proxy that sits between your applications and the model providers is a natural place to do the tagging, because every request already passes through it.
OpenLM’s Enterprise AI Gateway works this way. It attributes every token to a user, project or department, and per-token pricing set in SAM (software asset management) turns that usage into cost, ROI and chargeback. Token consumption can be mapped to cost centers automatically, and the usage data can be exported in the FOCUS 1.3 format to existing FinOps tools.
Whichever approach you choose, the test is the same: can you take last month’s AI spend and explain it by user, project and cost center, without a spreadsheet exercise?
Frequently asked questions
What is AI unit economics?
AI unit economics measures what AI costs per unit of business output, such as per ticket resolved, per query or per user. It turns raw token counts into numbers that finance and business owners can compare and act on.
What is the difference between showback and chargeback for AI?
Showback reports each team’s AI cost without moving money. Chargeback bills that cost to the team’s own budget or cost center. Showback is often the first step, and chargeback follows once the attribution is trusted.
How do you calculate cost per query for an LLM?
Divide the AI spend for a use case by the number of requests it handled in the same period, and quote the result per 1,000 queries if the figure is small. The spend should include every model call the use case makes.
How do you allocate AI costs to departments?
Capture the user, project and department for each request, apply a price per token to the tokens consumed, and add the results by cost center. Trying to rebuild this from a provider invoice at the end of the month rarely works, because the invoice does not show who made each request.
How do you measure AI ROI?
Compare the cost per outcome with the value of that outcome, such as the time saved on a resolved ticket. The value is usually an estimate, so agree it with finance in advance and report it as an assumption.
